Multi-task cascaded assessment of signal quality for long-term single-lead ECG monitoring. (May 2023)

• Record Type: Journal Article
• Authors: Liu, Sen; Zhong, Gaoyan; He, Jiacheng; Yang, Cuiwei
• Is Part Of: Biomedical signal processing and control. Volume 83(2023)
• Publication Date: 2023-05
• Subjects: Cascaded -- Signal quality -- Single-lead ECG -- Long-term monitoring -- Wearable devices; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2023.104674
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: The present three types of ECG signal quality can better accommodate follow-up analysis in practical applications. A novel multi-task cascaded method is proposed for single-lead ECG in long-term monitoring. Feature selection is performed to screen optimal task-specific combination. The proposed method shows excellent generalization on various datasets. Abstract: Background and objectives: In long-term electrocardiogram (ECG) monitoring, the use of wearable devices for signal acquisition inevitably introduces noise, which could reduce the diagnostic capability. In order to improve the efficiency of signal analysis and the accuracy of clinical diagnosis, it is necessary to automatically evaluate the signal quality of ambulatory ECG. In this study, a novel multi-task cascaded method is proposed for signal quality assessment. Methods: Based on single-lead ECG, various signal quality indices were derived to assess three classes of signal qualities (clear ECG ideal for full wave analysis, mildly contaminated ECG available for ventricular activity-based analysis, severely contaminated ECG difficult to be analyzed) through a two-step classification corresponding to different tasks. The optimized subsets of quality indices were selected with a criterion of maximal-mean-minimal-variance. Different machine learning models were then trained and tested on public datasets. Results: The results show that random forest using the optimal feature subsets had the best performance, with the overall classification accuracy of 93.76%, 95.13% and 99.61% on the three unknown test sets, respectively. Meanwhile, for the identification of acceptable and unacceptable signals in the first task, the AUC on the test sets reached 0.9967, 0.9959 and 1.0000, respectively. Conclusions: This paper presents a novel signal quality assessment framework, which could provide helpful feedback for ECG processing in specific application scenarios. It is expected to greatly promote accurate and efficient ECG-based diagnosis in long-term monitoring.